ProtoBind-Diff

ProtoBind-Diff generates small-molecule ligands for a protein target using nothing but that protein's amino acid sequence. Most target-conditioned generative chemistry models build molecules inside a resolved 3D binding pocket, tying them to structural datasets that are small and skewed: PDBbind holds fewer than 30,000 protein–ligand complexes, concentrated on well-studied targets and familiar chemotypes. Conditioning on sequence embeddings sidesteps that bottleneck, making any protein with a FASTA record a valid target.

Released as a bioRxiv preprint in June 2025 by Lukia Mistryukova, Vladimir Manuilov, Konstantin Avchaciov, and Peter Fedichev at Gero, a Singapore-based longevity biotech, the model frames ligand generation as masked discrete diffusion over SMILES tokens: a fully masked string is progressively unmasked over a fixed number of denoising steps, with protein context injected through cross-attention on frozen ESM-2 embeddings. Training draws on more than a million measured active protein–ligand pairs from BindingDB — roughly forty times the size of the structural corpora available to pocket-based methods.

The result never sees a 3D coordinate yet produces molecules that score well under structure-based evaluation. Benchmarked against pocket-conditioned generators including Pocket2Mol, PocketFlow, TargetDiff, and TamGen, ProtoBind-Diff achieves the highest enrichment factor when candidates are scored with Boltz-1, and its property distributions track known actives more closely overall.

#Key Features

• Sequence-only conditioning: Generation requires only the target's amino acid sequence, extending generative ligand design to orphan receptors, disordered proteins, and rapidly evolving pathogen targets where no pocket structure exists.
• Masked discrete diffusion over SMILES: Non-autoregressive denoising over a chemical token vocabulary allows the model to revise earlier decisions during sampling, which yields higher validity and diversity than the graph-diffusion variants the authors tried first.
• Emergent binding-site attention: Across 1,843 BioLiP-2 annotated proteins, a single cross-attention head predicts contact residues with a mean ROC-AUC of 0.716, despite no pocket or structural supervision during training.
• Robustness on data-sparse targets: Similarity clustering and SMILES augmentation improve generalization to targets with few recorded actives, where the model matches or exceeds structure-based baselines on Boltz-1 ipTM.
• Released weights and code: The checkpoint, tokenizer, and training pipeline are public under CC BY-NC 4.0 — free for research, closed to commercial use.

#Technical Details

The generator is a 12-layer transformer decoder with 8 attention heads and a hidden dimension of 1280, using rotary position embeddings and SUBS parameterization with a log-linear noise schedule. Protein conditioning comes from esm2_t33_650M_UR50D, whose 1280-dimensional per-residue embeddings are projected into cross-attention; the decoder width was chosen to match. Training used a learning rate of 5×10⁻⁵, dropout 0.1, and batch size 48. The dataset was built from the February 2025 BindingDB release of 3,010,313 measurements, filtered to sequences of 50–1,500 residues and ligands of 10–80 heavy atoms, with cytochrome P450 and albumin removed as promiscuous binders. A pair was labeled active when any of Ki, Kd, or EC50 fell below 1 µM, giving 1,167,809 training samples. Molecules within Tanimoto similarity 0.85 were clustered and resampled each epoch to reduce redundancy.

Inference generates 170-token SMILES strings over 250 denoising steps with nucleus sampling at p = 0.9 and the remdm-cap remasking scheme. Evaluation used 12 targets drawn from the intersection of CrossDocked2020 and BindingDB — six with over 1,000 training examples and six data-sparse ones — with 1,000 molecules generated per target. Under AutoDock Vina, ProtoBind-Diff's enrichment factor ranked below Pocket2Mol and PocketFlow, both trained on Vina-augmented poses; under Boltz-1 ipTM, which separated known actives from inactives more reliably, ProtoBind-Diff ranked first. Target-class separability remained weak in absolute terms: a k-NN classifier over UMAP-projected fingerprints of novel generated molecules reached only 0.23 accuracy, the best of any model tested.

#Applications

The practical draw is early-stage hit finding for targets that structure-based pipelines cannot reach — orphan GPCRs, intrinsically disordered proteins, fast-mutating pathogen proteins, and neglected-disease targets with no deposited complex. Because generation needs only a sequence embedding, medicinal chemistry teams can produce candidate sets for a new target within days and triage them with conventional downstream tooling: conformer generation, docking, or co-folding scorers. It also suits multi-target campaigns where enumerating pockets across a whole protein family would be prohibitive.

#Impact

ProtoBind-Diff is evidence that protein language model embeddings carry enough interaction signal to drive chemistry generation without explicit geometry, and that the resulting molecules survive structure-based scrutiny. The attention analysis is the sharper result: spatial priors emerge from sequence–ligand co-occurrence alone, though the model's head still trails a supervised ESM-2 linear probe (ROC-AUC 0.849) at binding-site detection. Substantive limitations remain: no 3D pose is produced, sequence embeddings omit pocket-level electrostatics and induced fit, generation skews toward the membrane receptors and enzymes that dominate BindingDB, and all validation is in silico — no generated compound has been synthesized or assayed. The work remains a preprint awaiting peer review, and the CC BY-NC 4.0 license rules out commercial deployment.